AI Agents or Automation? How to Choose the Right Approach

0:00There are very clear cases for automations and then clear cases for agents. and don't try to force an agent to do something that an automation can do easily. Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. If you're even loosely staying up to date in the AI space, you've likely heard of agents, AI automation, and some of the tools being used from Make to N8N to about 100 others. But the tooling and the hyped up terms are only like a small part of the equation.

0:38The important part is actually determining what the heck do you automate. But lucky for you, we got Marcus McGee, the founder of the AI Consulting Lab here to talk some AI. And we're actually going to show off some live examples today. So make sure you check that part out on YouTube. I've got a lot of good stuff coming. Marcus, welcome to the show. Thanks for having me. I appreciate it. Good to be here. Awesome. Yeah, this is going to be a good one. But let's start. I just want to start very basic. Sure. How do you define automation and agents? I think those are the hyped up terms right now. And I guess too, people understand I think what automation is, but why is it different with AI now?

1:15Yeah. Yeah. I think anyone that is in the space, whether they're doing automation or just AI in general, they're for sure jumping on the buzzword of AI agents. Everybody is using AI agents this year. and I think back in Q4 of last year, we had Salesforce and we had Microsoft and Google. All of them were saying AI agents were going to be the topic for 2025 and for sure it's here. So it's - Whether they manifested that into reality or I think it's a bit of both. Yeah, force manifestation, I think is a good view for that. Yeah, but any company really that's doing anything in the space, if this more likely that they're using the term just because it's getting more clicks, honestly.

1:56I think that's the big thing. But I did a video on this recently about AI agents versus automations. So it's good to think about it this way of an automation is just a sequence of steps, right? You program it ahead of time, and then it does the thing. There can be some AI components in there, individual modules like in Make, it's an AI module, and it does what you ask it to do. But it's still not really doesn't have much autonomy. But I like to say with an agent is think about it as a sports field, right? Maybe baseball field or football. they we give the team the tools the plays to run and then depending on how the defense is set up they can run the plays in different ways so same thing goes with agents right we can give it an overall objective right here's what you're supposed to do and here are all the tools that you have and here's how those tools function depending on how the situation comes in with the chat or however you're interacting use your tools to best do the job that you're supposed to do and here's your objective.

2:54That's I think an easy way for people to understand it. It doesn't really go like you're limited to that. So a true AI agent is autonomous and it can do things like opening AI's operator or Claude's computer use. That really is no bounds. It just has its objective to go and do. But the things like NADN agents and make agents, they operate within that boundary, right? So if you go take a football player, for example, in this example, and you put him on the middle of the street, it's not going to be the same thing, right? So he's not going to be able to use this tool appropriately and it's going to just be a total failure.

3:25Okay. What's the, oh my gosh, I'm like going back in time to that commercial. Is it Terry Crews, like the office linebacker? Yes. I had that like visual coming into my head of him in the office setting and just tackling somebody. But yeah, that's such a good example. It's so much, the one I lead with is the IVR and back in the day, like you had to call center. You had to define every scenario, if this, then that, every edge case. And if you didn't define it, you're just SOL, right? But with these systems now, these LLMs and agents, they're probabilistic. So they can quote unquote, think through a situation based on whatever their goal is and whatnot.

4:02But now I love the football example because it's about them being able to essentially observe whatever scenario they have going on. They have a goal, which in the football analogy, it's to go score, or maybe it's specific to the play they're running. And then they execute it with the tools they have, which is their teammates, football, things like that. But no, that's a good one. I like that. And it's, well, I watched a video the other day of this guy, another creator, smaller creator, but he was talking about differences in AI agents and automations. And it really stuck home because the message was it's bet there are clear use cases for automations and there are clear use cases for agents.

4:41And just because we have this new agent technology and it's cool and does all these things. Yes. Yes. It doesn't do. if you have a certain outcome that you want, you should stick to an automation. Myself, this is a prime candidate. Like I wanted to do, I have a blog creation and social content repurposing automation. And I made it make before agents were a thing. And then I really got fixated on trying to do it with an agent. And it just failed almost every single time. And I wasted a bunch of time trying to get it set up. Quick break in the pod. If you're listening to this podcast, chances are you've been thinking about how to actually use AI inside your business.

5:18And that's exactly why we built the AI Opportunity Finder. It's a free tool that helps you uncover high impact, tailored AI use cases based on your business, your goals, your pain points, and your industry. No fluff, no generic use cases, just real ideas that fit your business and the Ranked by ROI potential. It takes about three minutes to run and it's like having your own personal AI strategist for free. If you wanna try it for free, check out the link in the show notes or go to hatchworks.com backslash AI dash opportunity dash finder. So there are very clear cases for automations and then clear cases for agents.

5:53And don't try to force an agent to do something that an automation can do easily. Yeah, that's a great distinction. I think that's the shiny object syndrome. When anything new comes out, it's like, how can I use this cool new toy when it's complete overkill half the time? And then also with an LLM, the cost a lot of times can be minuscule, but if it's just a standard automation, it's a lot of times even cheaper than that. But it's a good segue though, because I think the core problem is folks see the shiny tool, the N8N or the make or name your automation tools. Oh, I want that. But they haven't actually thought through, okay, what do you want to automate?

6:33How does your process work today? and you do a lot of this work with your clients thinking through, okay, what do you automate? How do you automate? Where's the opportunity? Where's the pain? Where are the gaps? Talk, talk to me through what is your approach there? So folks listening can take some of these nuggets and maybe start to apply it in their business. So the framework that I kind of use for this, it approaches not only automations, but it's also really good for just manual AI assisted GPT stuff too. So yeah, let's take a step back. Let's go all the way back and because I think that's it.

7:05You got to get to the core. It's not, don't jump to the thing first, right? So when we look at putting AI in our business, if we're not using it right now, or we're just starting to use it, and we're getting familiar with it in our personal lives, and now we want to start using it in our business lives, to look at what are the things that we do, and I say this all the time, this is something that I repeat so many times to businesses that I talk to and people I work with, is look at the tasks that you're doing every single day. What are the ones that are taking 30 minutes, 45 minutes, or more that have a set framework.

7:36They do, you always have to do this step first and this one and this one. Look for those problems or look for those tasks and then look at the individual steps that are in those tasks. And don't try to look at it and say, all right, well, I do all of this to do one task and I asked ChatGPT to do it and it totally failed miserably. Yeah, because it's a complex task with a lot of steps and then you would probably have to go back and forth with it. It's not going to be able to do that. But if you take that and you break it down, Can it do three of the tasks for the overall five task total? It can do that.

8:10Okay, great. You've just taken out 60 % of the time that it would spend to take you to do all the menial work, the mundane work to get to those last two tasks. That's the first point, right? Figuring out how to break down your tasks into manageable chunks and figuring out which of those chunks can be handled by AI. That's the number one thing. Then you just - Pause there real quick. Cause I think, I feel like that sounds like such an easy thing to do. It's oh, this stuff I do all the time. Let's get that. But I feel like in my own experience, I'm thinking in my head, I don't know what those things are.

8:43So I feel like there's almost this, I was listening to his Alex Hermosi the other day. He was talking through something similar, but not necessarily automation related. He was just trying to figure out how to optimize his day. And he literally for a week, he broke his day out in the 15 minute increments. It would just jot down, okay, what did I do? What did I do? And it was beautiful as you can take that, feed that into whatever LLM you want and have it identify where the things are. But you almost have to do that like intentional bit to truly find out, okay, where is the repetitive work, the mundane tasks, the things that I really shouldn't have to do now that we have this awesome new tooling.

9:22Yeah. But so another good point of that. So going with the same thing of your Harmozi example, if you're not, because that's a skill. It's actually one of the things I have a, so in my corporate training that I do two days, an hour and a half each day, the first day is all about like intro to chat GPT, generative AI and how it works. Crash course of prompt engineering and kind of some of the limitations and do's and don'ts. And then the second day is about more advanced prompt engineering, but more importantly, how to build custom GPTs. And that section, segment particularly, talks about how to break down your tasks into this, exactly what we're talking about.

9:57And it's a skill, right? It is a very much a learned skill by understanding what the models are capable of. First and foremost, can it actually do something? I'm pretty good at it now to figure out. Somebody asked me to do something, break something down. I can look at the problem and say, all right, I know you need to do these things. And GPT is not going to be able to do those things. But they can do this and this, right? But if you're struggling with that, or maybe just, again, want to leverage the tool, go into one of the language models and say, here's the task. Here's what I normally do to do this task.

10:29Help me identify what are the areas of opportunity or how can I break this down further into tasks that a language model could do reliably every single time. That's a great way to do it, right? And then you can take it from there and go a step further and or maybe just not have your expectations so high. if we're looking at trying to cut down the total amount of tasks that we do in a set number maybe it's not that but maybe it even goes a level deeper like inception llm inception right can of these five tasks can chat gpt do 60 or 70 percent of each one of those tasks and then i only have to do the little bit afterwards so there's a couple of ways i think you can approach that situation the human in the loop there yeah yeah and then that carries over to your automations Right.

11:14So as soon as you figure out what those are, that's the basis for then putting it in your automation. Yeah. And I think too, the point you mentioned a second ago, it is a learned skill, but I think most things can be improved with an LLM and AI. But the key thing is you've got to break it down to the right molecular level in a sense, because if you're trying to have it do too much, it probably is going to fail. That's such an important piece. And then you add the elements of the AI's ability to reason. I think that adds a whole nother capability in a sense. And obviously we'll get to the automation and access to tools.

11:53And now I can actually do some things, but like the custom GPTs, I feel like that's such a great gateway into this. Cause it's almost like a lightweight proof of concepting tool to help automate a task. It's like this buddy that specialized at this thing that has context for what you're trying to do. And it can help you with it. Yeah. Like custom GPTs are probably the lowest hanging fruit for anybody that wants to adopt AI in their business. Because exactly just because of what you said, right? It is the low code, no code, really, solution to building an MVP that can deliver immediate return on investment.

12:32And if you look at that and define it as time save or more efficiency doing a specific type of project, that's your ROI. I have a client that she's on TikTok now. Now she's an immigration, she owns an immigration attorney firm in Arizona. And we worked with her at the beginning of last year to set up a system inside of Microsoft Copilot. But essentially it was more or less a set of custom GPTs. And their main task that they do is writing a draft declaration for immigrants to present for their court case. They told me that the process to write it, the actual writing process was around nine hours or so.

13:06And so we developed these custom GPTs that did a particular part of each one of their frameworks. So they had six parts to their framework for this. And I developed one for each one of those. And going down the line, putting in the transcript of the meeting with the client and the notes and things like that, in the hands of their most capable drafter, she was able to take that process down to 45 minutes. So nine hours to 45 minutes is a huge jump in time saving. And it's all just really a custom GPT, honestly. What you take here is I feel like when custom GPTs came out, I was just like, oh my God.

13:41This is the thing that's just going to change everything. I feel like it just did not become a major focus to continue building it out by open AI because they have so many other things. I feel like there's so much further they could take it, but I just feel like they haven't. I feel like it hasn't been a priority for them. Like as it is today is very functional, but, and I could be wrong, but I feel like they haven't added additional functionality to the custom GPTs past like the initial rollout. I would agree with that. I think that it's still serve a very valuable purpose to get people familiar with that process of building a specialized assistant with special knowledge.

14:24And then also the ability to connect third parties with the actions. I don't think I don't think that they really took it further than that but you could see a world where they almost go into the true autumn almost like encroaching in on the n8n space but then they have the operator too which is the browser tool I don't know it's just kind of interesting like strategically where they focus their time and their product roadmap and then there's curious too like projects do you play with those at all because I feel like those have some overlap in a sense, because you can add context to those. So just, I know we're going deep into chat GPT, but I like to nerd out on this sometimes, do you use those at all?

15:02And how do you differentiate between those and custom GPTs? Yeah. I think the biggest thing with open AI specifically is that they have a really bad, they have a problem with their product ecosystem and they just hired a new CEO for that. That it was actually, was it this week or last week they hired a new CEO for product specific. Did they really? I didn't know that. Yeah, it was the person used to be CEO of Facebook's video gaming and like the app, the Facebook app itself. So they brought her on in a CEO position. So it's two CEOs, but it's CEO of applications, I believe. I listened to this on one of the podcasts this morning, so it's still fresh in my mind.

15:43But I think that's the point. They know they have a problem with branding in terms of the model picker confusion, mess. they know that's 4.0, 0.3, 4.1. And that's okay. We've gone backwards in numbers, but it's a better model. Yeah, there was, again, another small person, a creator on Twitter, actually, X, however you want to phrase it. They, this was just like random to see this, but they were just tweeted, maybe OpenAI or one of the product managers or somebody at OpenAI and said, you guys have a problem with this. It's so confusing. The model picker is terrible. What do we use? And one of the management level executives responded to that tweet, and they were like, we know, we're working on it.

16:29So they're aware of the problem that they have an issue with that. Sam is too. I think either Sam or Greg Brockman talked about the model picker and how they want to make it easier. So they know they're working on it. But the projects and the custom GPTs, I think it's just by nature. They came out after each other. I think that honestly, I think that they should go together. I think that they should just roll those two in together because you can't change the models in custom GPTs. That's a big drawback. I think that's the biggest flaw to custom GPTs. And the limit on files to 20, I don't really get that, especially when you go into the developer platform and the assistance API can let you do 10 ,000 file attachments.

17:09Yeah. I don't know what the deal with that is either. and on the other side of that it does really good with your instructions and your knowledge but on projects doesn't do so good with the instructions and the knowledge but gives you the ability to choose the models right and yeah so just combine those together for me right now honestly i just use projects as a folder organization system so working with clients whatever else putting chats where i know if i had to go do new chats all the time or go back to them I'm just going to create a project folder and put it there. No instructions, very little.

17:42I might put some instructions there. I might put some knowledge, but for the most time, it's really just an organization system. Okay. Yeah. Sorry for the derailment there, everybody, but I had to get, I had to go deep on the opening. But yeah, let's actually do this. So we've talked a little bit about the automation. Let's just pull one up live. I think you got one in N8N. Talk us through it. I think it's just a good, simple example for people to visualize what this is actually doing. And then it, I, hopefully it'll get your wheels turning a bit in terms of the art of the possible in terms of what you can do.

18:14So like talk to me about this. I have no clue what I'm even looking at and I've never heard. And what do you call it? N8N or Natan? That's my first question. I call it N8N. What's the actual name for it? Yeah. I've heard people call it both. I've always called it N8N, but I've heard people call it Natan. Almost. It's like Nathan. I think the tech industry needs to leverage their tech better to come up with good names that people understand. That's one of the takeaways here. There you go. So this is just a pretty simple agent, right? This is a friend of mine. He's also here. Hopefully we'll see this video.

18:48But he asked me to put together an agent to help his company. They import charcoal. They're based in the UAE. company and they want to just basically take a look at their sales and keep track of how their sales are doing. So plugging it into their data with their reports and being able to get insights on that and forecasting. And he also said, I wanted to be able to do some competitor research. So what are my competitors doing? Like, how can I, what, what's the social monitoring of them and what are they doing on social and things, things like that. And of course, the ability to communicate that with external people, whether that's the partners or any clients or anything like that, having a mail module.

19:28So that's essentially what this is. I love NADN. I'm actually pretty relatively new to it. Like I had played around with it a little bit in the beginning or a Q4 last year, but it was pretty, pretty pricey comparative to Zapier and Make. So I didn't mess with it too much. And then your video, I'm pretty sure it was yours, learned, I learned that you can self-host it. And after that it was, all right, I'm off to the races now. Self-hosted, no limit on the number of workflows or executions you can do. So let's just start building. Yeah, it's just a simple chat here. You can embed this on your site, call it as a web look, whatever else.

20:01And it goes to the agent. You open the agent up and you can set your system assistant message just as if you would any other AI module in a fake automation or Zapier or things like that. And then you set your prompt. This is such an important step too. Like it's the whole concept of prompt engineering. I've gone back and forth where people call it a career. I think it's just a skill that people are going to have, but it's so critical. I think he gets overlooked a lot in the building process because this is where you can really fine tune and hone in what you want the thing to do. Yeah, absolutely.

20:33The system assistant message is the number one thing. And again, yeah, you're totally right on the prompt engineering stuff. People thought it was going to be the next career and then it turned into, it's important, very important, but as the - It's like typing and Googling and using Excel and whatnot. You totally have to be able to do that now. It's like a, it's just a basic necessity. And also too, the frontier model companies, Claude, they have their prompt generator. I love that. I use it so often. All that yesterday. I have not tried that yet. So in the console optimized. Yeah. In the console, you can create it from scratch or have it optimize it.

21:06If I have to come up with a prompt almost eight times out of 10, I'm going there first and just putting it in there and tweaking it from behind. But that's where the prompt engineering comes into play, right? because if you are have a specific goal, the, these prompt generators are not going to know your specific goal. So you have to go in and be able to diagnose, okay, I don't need this part. And let me change this one. And even then, once you've got something like this, you start testing it and you find that it's not doing what exactly what you want it to be. You have to figure out what part of the prompt do I need to change?

21:39And if I change things, does it impact the rest of the prompt and all this other stuff? So there's a ton of stuff. You got to go through that process. Yeah. It just allows there too. I know I keep going down rabbit holes, but this is another thing. If you're new or you just feel like you're not getting a lot of value in AI and whatever, Claude, doesn't matter. It's people think of prompting and using it like a new form of Googling, which it's not. It's completely different. I think the example you just gave, it's like these meta use cases of I'm actually talking with AI to help me create a better, more refined prompt that I'm going to use for AI.

22:16It's things like that. I use it so much just in the brainstorming to the outlining to whatever the next phase is that iterative approach. I think that's really important for people to start learning that skill, building that muscle. Yeah, absolutely. For sure. Because like, again, my proficiency came from just doing, man, I got access to chat GPT right around when it first came out. And I was writing a course for digital marketing to teach. I was putting it together an offer for that was based off high level to put it to mental health professionals. And I was combining all of my digital marketing knowledge into one to teach them how to market their business so they can use the tool effectively.

22:58ChatTPT came out and I just like everything. I just started running it through there. There was no guidelines on out of prompt, what you could and couldn't do. And I just started talking to it like it was a person. And can you do this? Can you do that? Okay, give me this. But actually no I don't like the way you wrote that write it in this way and change this information and it just through that process I started doing that for everything and I that's how I got to be so good at this whole industry right was just that process of doing every single task for myself and my clients in chat GPT and just there was nothing that it couldn't do in my mind until I totally exhausted all the ways to try to do it and then I figured out all right it can't do this part but I learned a lot along the way, right?

23:43Or just learn by doing, just go talk to it, figure out, and then ask follow-ups. How could I improve this prompt? Or what could be better on the backend? Did that this morning. Right before we jumped on here, I finished editing a video talking about using creator search insights and ChatGPT to grow your TikTok. And at the end, I wrote the prompt myself. And at the end of it, I was like, what else could I improve in that prompt? And it actually gave me two good suggestions that I did include in the new prompt. So definitely recommend doing that for sure, too. No, I did it for this podcast. We do a prep call before we do these.

24:14I fed the transcript into open AI. I gave it context for what I was doing. I say, Hey, create some structure around this podcast. Give me some high level questions to focus in on. It's just like little things like that. It just, yeah. All right. Let's get back to the automation. I was about to say the same thing. Yeah. The simple setup of this, you set your agent module, you put just a message telling it what its objective and goal is. And then also not only that telling it what tools it has access to. So Gmail, Google Sheets, perplexity, and the internal database. So it knows what these things are and how to use them.